A8gent

Deepgram Voice Agent API Pricing and Review (2026)

One WebSocket that bundles speech to text, a language model and text to speech at a single per-minute rate.

Reviewed by A8gent
Contents
  1. Verdict
  2. What is Deepgram Voice Agent API?
  3. Best for, and not for
  4. Strengths and limits
  5. Pricing
  6. What people actually build with the Deepgram Voice Agent API
  7. What the Deepgram Voice Agent API actually costs at small-business volume
  8. The gotcha: the model tier doubles the rate, and silence is billed
  9. What breaks in production, and how you find out
  10. What it connects to that you already run
  11. Deepgram Voice Agent API alternatives: what to pick instead, and when
  12. Migration and lock-in: can you get your work out?
  13. FAQ
VerdictIs Deepgram Voice Agent API worth it?

Deepgram's Voice Agent API is the simplest honest price in voice infrastructure: $0.075 a minute on a Standard model, $0.163 on an Advanced one, with speech to text, the model and text to speech all inside that number. At our reference volume of 4,000 minutes a month that is $300 on Standard, and about $353 once Twilio carries the calls. It is worth it if you have a developer and want one bill for the AI layers. It is the wrong purchase if you wanted a product: there is no phone number, no builder and no call dashboard, and the one decision that doubles your bill, the model tier, is a line in a config file.

What is Deepgram Voice Agent API?

The Deepgram Voice Agent API is a single WebSocket connection that runs the whole speech pipeline for a voice agent: it listens with Deepgram's speech to text, thinks with a language model you pick from a managed list or bring yourself, and speaks with text to speech. Deepgram bills the three together as one per-minute rate rather than as separate components. It is an API for developers, not a call-centre product: there is no phone number, no visual builder and no dashboard of call flows, and Deepgram's own telephony guide bridges calls through a Twilio number and a server you write.

Best for

Developers who want one bundled per-minute rate for the speech and model layers and are happy to write the telephony bridge themselves.

Not for

A business with no developer that wants a phone number, a builder and a working receptionist by Friday.

Strengths

  • One published per-minute rate covers speech to text, the language model and text to speech, so there is no separate model bill on managed models
  • Pay as you go has no minimum and no expiry, and the $200 free credit covers about 2,666 Standard minutes
  • Bring your own model, your own voice, or both, and the rate drops to as little as $0.050 a minute
  • 45 concurrent Voice Agent connections on Pay as you go without a sales conversation
  • Managed model list spans OpenAI, Anthropic, Google, NVIDIA and Groq, with each model's pricing tier published in the docs
  • Function calling lets the agent call your own APIs and tools in the middle of a conversation

Limitations

  • No phone numbers or carrier: telephony is a separate bill and a bridge server you build and host
  • Choosing an Advanced-tier model more than doubles the rate, from $0.075 to $0.163 a minute
  • Billing is on WebSocket connection time, so silence, hold time and a connection left open all cost the full rate
  • Sessions close automatically after 2 hours and must be restarted with the earlier turns replayed
  • The Growth plan is a prepaid annual commitment starting at $4,000, and overage on it bills at the Growth rate plus 10 percent
  • No visual builder, call analytics suite or non-developer interface

Deepgram Voice Agent API pricing

Deepgram bills the Voice Agent API per minute of WebSocket connection time. The rate depends on the language model tier (Standard or Advanced) and falls if you bring your own model, your own text to speech, or both. Pay as you go has no minimum; Growth is prepaid annual credit starting at $4,000 a year with lower rates; Enterprise is custom. Telephony is not included at any tier.

Deepgram Voice Agent API pricing tiers
PlanPriceWhat you get
Pay as you go, Standard$0.075/min$4.50 an hour. Deepgram speech to text, a Standard-tier managed model (for example gpt-4o-mini, claude-haiku-4-5 or gemini-2.5-flash) and Deepgram text to speech. No minimum, no expiry, $200 free credit, up to 45 concurrent connections
Pay as you go, Advanced$0.163/min$9.78 an hour. Same pipeline with an Advanced-tier managed model (for example gpt-4o, gpt-4.1, claude-sonnet-4-5 or gemini-3-pro-preview). $0.122 a minute if you bring your own text to speech
Bring your own model$0.065/minListed as Custom, BYO LLM. You pay your model provider separately. Bringing your own text to speech as well takes it to $0.050 a minute; Standard with your own text to speech only is $0.065
GrowthFrom $4,000/yrPrepaid annual credits. Standard $0.068, Advanced $0.146, BYO LLM $0.059, BYO LLM and TTS $0.041 a minute. Up to 60 concurrent connections in North America, 45 elsewhere. Overage bills at the Growth rate plus 10 percent, weekly
EnterpriseCustomCustom pricing for large volumes, data or deployment requirements and support needs. Concurrency starts at 100 connections

Pricing reflects public plans as of November 17, 2026 and can change. Check Deepgram Voice Agent API for the latest.

What people actually build with the Deepgram Voice Agent API

The Voice Agent API is a building block, so what gets built on it is whatever a developer was going to build anyway, minus the work of wiring three vendors together. The question it answers is narrow and useful: can I get speech to text, a model and a voice from one connection and one invoice. For these builds the answer is yes.

The builds that fit

An inbound phone line for a service business. A Twilio number receives the call, a small server bridges Twilio's audio stream to Deepgram's WebSocket, and the agent answers questions and books through function calls to your own scheduling API. This is the reference architecture in Deepgram's own telephony guide, and at 1,000 four-minute calls a month it is our reference workload.

A voice interface inside a web or mobile app. No carrier is involved at all, so the Deepgram rate is the whole AI bill. Deepgram ships a browser agent SDK for this, and it is the cheapest shape the product takes because there is no telephony line to add.

A voice front end for an agent you already have. If your text agent already runs on your own model endpoint, the bring-your-own-model tier at $0.065 a minute gives it ears and a voice without moving the reasoning anywhere.

An internal tool where staff speak rather than type: a warehouse pick confirmation, a field engineer logging a job hands-free. Short sessions, structured answers, a Standard-tier model, and a bill that is easy to predict because it is minutes times one rate.

A multi-agent handoff, where a triage agent passes the caller to a specialist agent by context or intent. Deepgram documents this pattern directly.

What it is genuinely bad at

Being a product. There is no phone number to buy, no visual flow editor, no call recordings library and no analytics for a manager to read. Every one of those is something you build, buy elsewhere or go without. A small business comparing this with a self-serve voice platform is comparing a component with a finished appliance, and the lower per-minute rate reflects exactly that difference.

It is also a poor fit for very long sessions. A session closes automatically after 2 hours; the server warns 5 minutes ahead and you must open a new session and replay the earlier turns as context.

Bottom line: Build on the Deepgram Voice Agent API when you have a developer and want the three AI layers on one bill; buy a voice platform instead when you need a phone number and a builder.

What the Deepgram Voice Agent API actually costs at small-business volume

Start with the published rates, read off deepgram.com/pricing on 17 November 2026. On Pay as you go, a Standard configuration is $0.075 a minute ($4.50 an hour) and an Advanced configuration is $0.163 a minute ($9.78 an hour). If you bring your own language model the rate is $0.065; bring your own text to speech as well and it is $0.050. Standard with only your own text to speech is $0.065, and Advanced with your own text to speech is $0.122.

At our reference workload of 1,000 inbound calls averaging 4 minutes, the arithmetic is one multiplication. 4,000 minutes times $0.075 is $300 a month on Standard. 4,000 times $0.163 is $652 on Advanced. 4,000 times $0.065 is $260 with your own model, plus whatever that model costs you. 4,000 times $0.050 is $200 with your own model and your own voice, plus both of those bills.

The free credit is real and does not expire. $200 divided by $0.075 is 2,666 minutes of Standard conversation, or about 44 hours. On Advanced the same $200 buys 1,226 minutes. That is enough to build and test an agent properly before a card is charged.

Then add the line Deepgram does not sell. Telephony is not part of the Voice Agent API. Deepgram's telephony guide uses a Twilio number and Twilio Media Streams. On Twilio's published US rates, local inbound is $0.0085 a minute, a local number is $1.15 a month and Media Streams is $0.0044 a minute. At 4,000 minutes that is $34.00 plus $1.15 plus $17.60, so $52.75. The all-in figure on Standard is therefore $300 plus $52.75, which is $352.75 a month, before the small server that bridges the two.

Deepgram Voice Agent API at 1,000 inbound calls a month averaging 4 minutes (4,000 minutes). Rates from deepgram.com/pricing and twilio.com/en-us/voice/pricing/us on 17 November 2026.
ConfigurationPay as you go rateAt 4,000 minutesWhat is still billed elsewhere
Standard (managed Standard-tier model)$0.075/min$300.00Telephony
Standard, bring your own text to speech$0.065/min$260.00Telephony, your voice provider
Custom, bring your own model$0.065/min$260.00Telephony, your model provider
Custom, bring your own model and text to speech$0.050/min$200.00Telephony, model and voice providers
Advanced (managed Advanced-tier model)$0.163/min$652.00Telephony
Advanced, bring your own text to speech$0.122/min$488.00Telephony, your voice provider
Twilio local inbound, number and Media Streams$0.0085/min + $1.15/mo + $0.0044/min$52.75Billed by Twilio
All-in, Standard plus Twilio-$352.75Hosting for the bridge server

When the Growth plan is cheaper, with the break-even shown

Growth is prepaid annual credit starting at $4,000 a year, which is $333.33 a month, and it lowers the Standard rate from $0.075 to $0.068. The trap is assuming a lower rate means a lower bill. At 4,000 minutes a month you would use 4,000 times $0.068, which is $272 of credit, against a commitment of $333.33. Pay as you go at the same volume is $300. Growth costs you $33 a month more.

The break-even is where Pay as you go spend equals the commitment: $333.33 divided by $0.075 is 4,444 minutes a month. Below that, stay on Pay as you go. Between 4,444 and 4,902 minutes ($333.33 divided by $0.068) you are paying the flat commitment and it is still cheaper than Pay as you go. Above 4,902 minutes you are into overage, which Deepgram bills at the Growth rate plus 10 percent, so $0.0748 a minute on Standard, a fraction under the Pay as you go rate. Growth credit is spendable across Deepgram's other products too, so a business that also transcribes recordings reaches the commitment sooner.

Bottom line: At 4,000 minutes a month the Deepgram Voice Agent API is $300 on a Standard model and $652 on an Advanced one, and the Growth plan only beats Pay as you go above 4,444 minutes a month.

The gotcha: the model tier doubles the rate, and silence is billed

There are two things on the pricing page that decide your bill and neither is a plan name. The first is the model tier. Deepgram's documentation lists every managed model against a pricing tier, and the difference between them is not a small premium: Standard is $0.075 a minute and Advanced is $0.163, which is 2.17 times the price for the same call. On the list we read, gpt-4o-mini, gpt-5-mini, claude-haiku-4-5 and gemini-2.5-flash are Standard; gpt-4o, gpt-4.1, gpt-5, claude-sonnet-4-5 and gemini-3-pro-preview are Advanced. Changing one string in the agent settings moves 4,000 minutes from $300 to $652.

That is worth deciding deliberately. Most phone conversations a small business handles are short, structured and tool-driven, and a Standard-tier model is the right default for them. If a developer reaches for the biggest model out of habit, the invoice more than doubles and nobody in the business will know why, because the plan did not change.

The second is the billing unit. The pricing page states that the Voice Agent API is billed on WebSocket connection time. Not speech time, not model tokens, not characters spoken: the time the connection is open. A caller on hold while your API looks something up is billed. A pause while they find their booking reference is billed. A connection your server forgets to close after the caller hangs up is billed until it does close, or until the 2-hour session limit closes it for you. One leaked connection that runs to the limit costs 120 minutes times $0.075, which is $9, for a call that may have lasted ninety seconds.

The practical fix is unglamorous. Close the Deepgram socket the moment the carrier reports the call ended, set your own maximum call length well under two hours, and reconcile Deepgram minutes against carrier minutes monthly. If the Deepgram figure is noticeably higher, connections are leaking.

A smaller one for Growth customers: overage is charged weekly, on Mondays, at the Growth rate plus 10 percent, and the pricing page says requests return a 402 once credits run out if no card is on file. A prepaid plan that runs dry therefore fails closed, on live calls, unless billing is set up for overage in advance.

Bottom line: Deepgram bills connection time, not talk time, and an Advanced-tier model costs 2.17 times a Standard one, so the two settings that decide the bill are the model string and how reliably you close the socket.

What breaks in production, and how you find out

The first limit you meet is concurrency. Deepgram's rate limits page gives Pay as you go up to 45 concurrent Voice Agent connections, Growth up to 60 in North America and 45 in Europe, Australia and India, and Enterprise a starting point of 100. For a small business 45 simultaneous calls is generous. Two details matter more than the headline number. Limits apply per project, and the same page says secondary projects on a self-serve account are limited to a single concurrent stream. A team that creates a second project for staging or for a second client and points production traffic at it will find the second caller cannot connect.

The second is the 2-hour session ceiling. The server sends a warning 5 minutes before, then closes the session with an error. For a phone agent this almost never matters. For an always-listening kiosk or an in-app companion it is a hard requirement to handle: open a new session and replay prior turns as context, or the user experiences it as the agent forgetting everything mid-sentence.

The third is the part Deepgram does not run: your bridge. In the documented telephony architecture your server holds two WebSocket connections per call, one to Twilio and one to Deepgram, decoding and re-encoding audio between them. That server is the component most likely to fail and the one nobody monitors, because it sits between two vendors who each report themselves healthy. Give it health checks, alert on calls that connect at the carrier and never open a Deepgram session, and validate the carrier's request signature so strangers cannot open sessions on your account.

The generic voice failures apply as they do everywhere: mishearing names, postcodes and reference numbers, talking over a caller who pauses mid-thought, and a human escalation path that was never tested under load. Test with real recordings of your own customers rather than a quiet office, and have the agent read critical details back.

Compliance, and what we could not verify

The pricing page describes Enterprise as covering data or deployment requirements, and the documentation offers a self-hosted option. We did not find a published price, a HIPAA statement or a data-retention term for the Voice Agent API on the pages we read on 17 November 2026, so we are not reporting one. If you handle health or payment information on calls, get the answer in writing from Deepgram before any of it reaches a live line, and remember that call recording consent is your obligation whichever vendor carries the audio.

Bottom line: The failure most likely to hurt on Deepgram is in the server you wrote between the carrier and the WebSocket, so monitor the bridge, not just the two vendors on either side of it.

What it connects to that you already run

Be clear about the model: there are no connectors here. The Voice Agent API integrates with your business through function calling, which means the agent decides mid-conversation to call a function, your code runs it against your calendar, CRM or booking system, and the result goes back into the conversation. Anything with an API is reachable. Nothing is reachable without code.

On the model side the choice is wide. The managed list covers OpenAI, Anthropic, Google, NVIDIA and Groq, and Deepgram hosts early-access models of its own. For bring your own, the documentation says any model string is accepted, and an OpenAI-compatible endpoint or gateway can be used by pointing the provider settings at it, which includes Azure OpenAI and Amazon Bedrock where the endpoint behaves like OpenAI's chat completions API. If your company already has a negotiated model contract, that is the route that uses it.

On the voice side, Deepgram's own text to speech is the default, and third-party voices are documented separately for teams that want a specific sound, at the bring-your-own text to speech rate.

For telephony the documented partner is Twilio, with the call arriving as a media stream. Nothing stops you using another carrier that can stream call audio to a WebSocket, but you will be writing that bridge without a reference implementation. For everything after the call, the sensible pattern is the one the rest of this directory recommends: have the agent post a structured summary to a webhook and let an automation platform such as n8n or Make.com update the systems of record, so the voice code stays small.

Bottom line: Deepgram connects to your business through function calls you write, and to almost any language model through an OpenAI-compatible endpoint, with no connector catalogue in between.

Deepgram Voice Agent API alternatives: what to pick instead, and when

The honest sorting question is how much of the stack you want to own. Deepgram sits in the middle: more bundled than a framework, far less finished than a platform.

Pick Twilio ConversationRelay if you are already on Twilio and want the speech layers managed inside the call itself, with no bridge server to run. It is $0.07 a minute on Twilio's price list, you bring the language model, and the call minutes are billed on top. Pick Cartesia Line if you want a managed runtime for the agent code as well: $0.03 a minute for the call, with the model and telephony billed separately.

Pick ElevenLabs Agents if the voice itself is what sells the call. Pick Retell AI or Bland AI if you want a self-serve platform with a phone number and a builder and do not want to write a bridge. Pick Vapi if you want a platform but with your own provider accounts underneath. And pick LiveKit Agents if you want to own the whole pipeline in code and treat Deepgram as just the speech to text inside it. The figures for those five in the table below are from our own August 2026 reviews and were not re-read today, so follow the links for the dated detail.

Alternatives at 4,000 minutes a month. Deepgram, Twilio and Cartesia rates read on 17 November 2026; the other rows cite this directory's August 2026 reviews.
PlatformBilling unitAt 4,000 minutesWho it suitsDoes better than Deepgram
Deepgram Voice Agent APIPer minute of connection, model included on managed tiers$300 Standard, $652 Advanced, before telephonyDevelopers who want one AI billModel, speech and voice in one rate
Twilio ConversationRelayPer minute, plus call minutes; model is yours$280 plus $35.15 of local inbound callingTeams already on TwilioNo bridge server; carrier and speech from one vendor
Cartesia LinePer minute, plus telephony and model$120 plus $56 telephony, before the modelDevelopers who want hosted agent codeLower published call rate; managed runtime
ElevenLabs AgentsSubscription minutes at a flat effective rate$320 at $0.08 a minute, model and telephony separate (August review)Brands where the voice is the productVoice quality
Retell AIItemised per-minute componentsAbout $354 on a lean configuration (August review)Teams wanting a platform with a builderPhone numbers, builder and call tooling included
VapiPlatform fee per minute, providers at cost$0.05 a minute platform fee before providers (August review)Teams holding their own provider accountsA finished platform over your own keys
LiveKit AgentsOpen-source framework; you pay the providersDepends on the providers you assembleEngineers who want the whole pipeline in codeTotal control of every layer

Bottom line: If you do not want to write and run a telephony bridge, a self-serve voice platform will cost more per minute than Deepgram and less overall.

Migration and lock-in: can you get your work out?

Lock-in here is light, and the reason is structural: Deepgram holds very little of your work. The agent's behaviour is a settings message your code sends when the connection opens, containing the prompt, the model choice, the voice and the function definitions. That configuration lives in your repository, not in a vendor dashboard. The functions themselves are your code calling your systems. The phone number belongs to your carrier account.

Pay as you go has no minimum and no expiry, so leaving is a matter of pointing your bridge at something else. Growth is different: it is prepaid annual credit, so the lock-in is whatever balance you have not spent. Size the first year's commitment against measured usage rather than a forecast.

What you would rebuild on a move is the protocol layer: the code that speaks Deepgram's WebSocket messages, handles its events and feeds function results back. That is days of work for a developer rather than weeks, and it is smaller still if you isolate it behind an interface from the start. What does not port is tuning: endpointing behaviour, how quickly the agent yields when interrupted, and the specific voice. Callers notice a changed voice more than any other change you can make.

The durable asset is the conversation design: what the agent says, what it must never say, how it confirms a spelling and when it hands over to a human. Keep that in a document outside any platform. The bring-your-own-model tier is also a quiet hedge: if the model already runs on your own endpoint, the reasoning layer moves with you untouched and only the speech layer changes.

Bottom line: On Pay as you go the only things Deepgram holds are minutes already used, so the exit cost is rewriting the WebSocket protocol code and re-tuning the voice.

Deepgram Voice Agent API FAQ

How much does the Deepgram Voice Agent API cost per minute?

On Pay as you go it is $0.075 a minute with a Standard-tier managed model and $0.163 a minute with an Advanced-tier one, as published on 17 November 2026. Bringing your own language model drops it to $0.065, and bringing your own model and text to speech drops it to $0.050. Growth plan rates are $0.068 Standard and $0.146 Advanced.

Does the per-minute rate include the language model?

Yes on the Standard and Advanced configurations: Deepgram's speech to text, a managed model and text to speech are all inside the one rate. On the bring-your-own-model configuration it does not, and you pay your model provider separately.

What is the difference between Standard and Advanced?

The language model. Deepgram's docs assign each managed model a pricing tier. Smaller models such as gpt-4o-mini, claude-haiku-4-5 and gemini-2.5-flash are Standard at $0.075 a minute; larger ones such as gpt-4o, claude-sonnet-4-5 and gemini-3-pro-preview are Advanced at $0.163.

How is usage measured?

By WebSocket connection time, according to the pricing page. Silence, hold time and any connection left open after a call ends are billed at the full rate, so close the connection as soon as the call finishes.

Does Deepgram provide a phone number?

No. The Voice Agent API is the speech and model pipeline only. Deepgram's telephony guide uses a Twilio number and a server you run to bridge the call audio. On Twilio's US rates that adds $52.75 at 4,000 local inbound minutes, including the number and Media Streams.

How far does the $200 free credit go?

$200 divided by $0.075 is 2,666 minutes of Standard conversation, about 44 hours. On Advanced it is 1,226 minutes. The credit does not expire and no card is required to start.

When is the Growth plan worth it?

Growth is prepaid annual credit from $4,000 a year, or $333.33 a month. Pay as you go on Standard reaches that spend at 4,444 minutes a month, so below that volume Pay as you go is cheaper unless you also spend the credit on other Deepgram products.

How many calls can run at once?

Up to 45 concurrent Voice Agent connections on Pay as you go, up to 60 in North America on Growth, and from 100 on Enterprise. Limits are per project, and secondary projects on a self-serve account are limited to one concurrent stream.

Is there a session length limit?

Yes. Sessions close automatically after 2 hours, with a warning 5 minutes before. To continue you open a new session and pass the earlier turns in as context.

Looking at alternatives? Deepgram sells the speech pipeline and nothing around it, so the booking, the CRM record and the follow-up message after the call are still yours to build. Autonoly is the opposite shape: a managed automation service for the work that happens after a conversation. If you have a developer and want a voice agent, Deepgram is a clean foundation. If you have a task and no developer, start with the workflow rather than the WebSocket. See the Autonoly review.